Artificial Intelligence-Driven Precision Agriculture: A Multi-Scale Framework for Climate-Resilient and Sustainable Food Production Systems

Abstract

This study developed a hybrid CNN-LSTM model utilizing multi-source data including satellite imagery, weather parameters, soil characteristics, and historical yield records from 2018-2024. The framework incorporates spatial-temporal feature extraction, attention mechanisms, and ensemble learning strategies. Model performance was evaluated against Random Forest, XGBoost, and Transformer baselines using cross-validation across diverse agro-climatic zones. The proposed CNN-LSTM hybrid achieved superior predictive accuracy (Rš = 0.946, RMSE = 0.342 ton/ha) compared to conventional approaches. Feature importance analysis identified soil moisture (SHAP = 0.184), temperature anomalies (SHAP = 0.156), and nitrogen levels (SHAP = 0.142) as primary yield determinants. The framework demonstrated robust generalization across climatic gradients with 89.3% accuracy in extreme weather scenarios. The AI-driven precision agriculture framework offers a scalable solution for sustainable intensification, enabling data-driven decision support for farmers while reducing environmental footprint. Integration of explainable AI techniques enhances trust and facilitates practical adoption in diverse agricultural contexts.